Impact of artifact reduction using generative adversarial networks on diagnostic accuracy in cone-beam computed tomography

To evaluate the impact of artifact reduction using a generative adversarial network (GAN) on the diagnostic performance for vertical root fracture (VRF) detection in cone-beam computed tomography (CBCT) images. A human mandible covered with Mix-D material and 40 extracted mandibular single-rooted teeth were used. All teeth underwent endodontic instrumentation, and VRF was induced in half of them. The teeth were then inserted into the canine and second premolar sockets on the left side of the image phantom. Scans were acquired in the CBCT unit CS 9300 (Carestream, Rochester, NY, USA) adjusted to a field of view of 5 × 5 cm, 0.09 mm voxel size, and two exposure protocols (1.100 mAs, 90 kVp, and 7.13 mGy cm 2 ; 2.24 mAs, 70 kVp, and 0.86 mGy cm 2 ). To simulate different clinical conditions, each root canal received one of the following materials: a fiberglass post, an R25 gutta-percha cone, or a cobalt-chromium post. Additionally, two titanium implants were alternately placed in the empty sockets of the left first molar and right mandibular canine, and additional scans were performed. A GAN-based neural network was developed according to the literature and adapted for training, validation, and testing using the CBCT dataset. Four observers evaluated the images and indicated the presence of VRF using a five-point confidence scale. The area under the receiving operating characteristic curve (AUC), sensitivity, and specificity were calculated and compared using analysis of variance (α=0.05). Overall, GAN-corrected images showed significantly higher AUC, sensitivity, and specificity (p<0.05) compared to the original images. The developed neural network demonstrated high performance in enhancing the diagnostic accuracy for VRF in artifact-affected CBCT images.